fix(detect): honor nested .gitignore/.graphifyignore files below the scan root

[Mohak-Agrawal]

Summary

Closes #1206.

detect() only read .gitignore/.graphifyignore files in the scan root and its ancestor directories (up to the nearest VCS root), loaded once before the walk began. A .gitignore sitting in a descendant directory — e.g. vendor/sub/.gitignore — was never read at all, so files/dirs it was meant to exclude leaked into the graph. Real git (and every other gitignore-aware tool) honors .gitignore at every directory level, not just the ancestor chain.

• Extracted the per-directory read+parse logic (previously inlined in _load_graphifyignore's ancestor loop) into a shared _load_dir_own_ignore() helper — no behavior change to the existing ancestor-chain path.
• detect()'s os.walk loop now calls _load_dir_own_ignore(dp) for every directory it visits (except the scan root, already covered), appending patterns before that directory's children are pruned — so a nested ignore file governs its own subtree with git's usual closer-file-wins precedence.

Test plan

• 3 new regression tests in tests/test_detect.py: nested file exclude, nested directory prune (walk never descends into it), nested ! negation overriding a broader root-level rule
• All 6 pre-existing .graphifyignore/ancestor-chain tests re-verified — no regressions
• Manually exercised all 9 tests end-to-end against real temp directories (this sandbox's Python has no pytest installed) — all pass; would appreciate a CI run to confirm under the project's actual test runner

🤖 Generated with Claude Code

[safishamsi]

Landed on v8 (8a5287a, your commit preserved) plus a test follow-up (7ca43ab) — ships in 0.9.15. Thanks @Mohak-Agrawal. The fix is correct and composes cleanly with the .git/info/exclude handling from #1810: nested patterns are read during the walk and anchored to their directory, so a nearer ! re-include still wins over both a root .gitignore and info/exclude (verified). I tightened the negation test to use .py files (deterministic code bucket instead of a fuzzy document/unclassified union) and added a composition test pinning the three-source precedence.